A comprehensive investigation of machine learning models for estimating
daily snow water equivalent over the Western U.S.
Abstract
Substantial progress on machine learning (ML) models and graphical
processing units (GPUs) has stimulated emerging research in applications
of ML to earth science. As snow is a vital component of the global
hydroclimate system, precise snowpack prediction is of considerable
value for science and society. In this work, we have trained three
different ML models (LSTM, CNN and Attention) to predict daily snow
water equivalent (SWE) with both dynamic and static features in the
Western Contiguous United States from Snow Telemetry (SNOTEL)
observations. Dynamic features include precipitation, minimum and
maximum temperature, minimum and maximum relative humidity, specific
humidity, solar radiation and wind velocity. Static features are
latitude, longitude, elevation, diurnal anisotropic heating (DAH) index
and topographic radiative aspect (TRASP) index. This choice of features
allows us to produce high-resolution maps of regional SWE for a given
set of input meteorological conditions. The importance and the
sensitivity of input variables will be tested by several explainable AI
methods including feature permutation and integrated gradient. The
ML-based dataset is further up-sampled and compared with the 4km gridded
SWE dataset from the National Snow & Ice Data Center (NSIDC), which is
from a physical-based model. Future SWE estimates are also produced
under climate conditions projected by CMIP class models, along with
associated uncertainty estimates based on our sensitivity analysis. The
ML models are demonstrated to be a fast and accurate method of producing
high-resolution SWE estimates with minimal computing power.